Abstract Achieving a harmonious balance between fairness and predictive accuracy is critical for trustworthy AI. We present a component-level causal reasoning framework that identifies and mitigates bias early in the pipeline—during imputation, encoding, and sampling—thereby preserving overall model utility. Our approach introduces unified Statistical, Causal, and Counterfactual metrics that quantify disparities alongside accuracy and F1 scores at each stage. By employing adaptive thresholds, we ensure fairness improvements do not compromise accuracy. This targeted mitigation uncovers root causes tied to sensitive and correlated features, applies interventions precisely where needed, and prevents downstream distortions. Evaluated on five benchmark datasets (Adult Census Income, German Credit, COMPAS, Bank Marketing, Titanic), our method maintains or enhances accuracy while substantially reducing bias, demonstrating a practical path to balanced, equitable AI. These findings lay the groundwork for the BIASGUARD framework—a step-by-step blueprint for integrating fairness-accuracy auditing and adaptive mitigation in production pipelines.
Rana et al. (Mon,) studied this question.